AI Paper Atlas

Research map / Large language models / Language, safety and interpretability

Scaling with Collapse: Efficient and Predictable Training of LLM Families

Shane Bergsma, Bin Claire Zhang, Nolan Simran Dey, Shaheer Muhammad, Gurpreet Gosal, Joel Hestness

ICLR 2026 · Large language models › Language, safety and interpretability · approach: LLM unlearning and benchmarks

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
match LR schedule, TPP and AdamW timescale tau across model sizesmuP LLM families · SlimPajama, 111M-3.3Bnormalized training loss curve collapseImproved
mis-scale tau across sizes (Llama-2 style families) (ablation)LLM families · Llama-2 curvesnormalized loss curve collapseHurt
increase tau (AdamW timescale) (ablation)610M, 80 TPP · SlimPajamaTLC shapeMixed
increase TPP at fixed tau (ablation)111M and 610M · SlimPajamaTLC shapeImproved
use CompleteP instead of muP (ablation)Celerity · Celerity pretrainingefficiency and reliabilityImproved
emphasize educational, math and coding data instead of general SlimPajama (ablation)Celerity · Celerity pretrainingdownstream qualityImproved
monitor collapse residuals against a small-model referenceCelerity 1.8B, 234 TPP run · Training diagnosticstime to detect divergence: divergence pinpointed near 60% vs raw loss rise only after 90%Improved
fix tau (adjust weight decay) rather than fixed weight decay in batch size sweeps (ablation)Batch size sweeps · Hyperparameter tuningcurve ordering for early stoppingImproved
fit b and q as power laws in tau and TPP instead of fixed values (ablation)Normalized TLC surrogate model · Surrogate fit, 111M dataMAE: MAE reduced by two-thirdsImproved
fit b and q jointly on both tau and TPP (ablation)Normalized TLC surrogate model · Surrogate fitMAENo clear effect
use predicted-best (collapse-aligned) early stopping vs current-bestWeight decay sweeps · 1.7B/20TPP and 3.3B/30TPPfinal loss gap of chosen setting: negligible gap after 30% and 10% of trainingImproved
use current-best selection at stop point instead of predicted-best (ablation)Weight decay sweeps · 1.7B/20TPP vs 3.3B/30TPPfinal loss gapMixed

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy